Strip raw inline JSON tool calls from local LLM display path (parity with cloud path)

Author: Brendan Gray

main/agenticChat.js

@@ -1107,10 +1107,11 @@ function register(ctx) {
 
       fullResponseText += responseText;
 
-      // Strip tool fences from display copy
+      // Strip tool fences and raw inline JSON tool calls from display copy
       let displayChunk = responseText
         .replace(/\n?```(?:json|tool_call|tool)\b[\s\S]*?```\n?/g, '')
         .replace(/\n?```(?:json|tool_call|tool)\b[\s\S]*$/g, '')
+        .replace(/\[?\s*\{\s*"(?:tool|name)"\s*:\s*"[^"]*"[\s\S]*?\}\s*\]?/g, '')
         .replace(/\n{3,}/g, '\n\n');
       if (continuationCount > 0) {
         displayChunk = displayChunk.replace(/\[(?:Continue your response|You were generating a tool call)[\s\S]*?\]/gi, '');
@@ -1576,6 +1577,8 @@ function register(ctx) {
     cleanResponse = cleanResponse.replace(/\n?```(?:json|tool_call|tool)\b[\s\S]*?```\n?/g, '');
     cleanResponse = cleanResponse.replace(/\n?```(?:json|tool_call|tool)\b[\s\S]*/g, '');
     cleanResponse = cleanResponse.replace(/^\s*```\s*$/gm, '');
+    // Strip raw inline JSON tool calls (same regex the cloud path uses)
+    cleanResponse = cleanResponse.replace(/\[?\s*\{\s*"(?:tool|name)"\s*:\s*"[^"]*"[\s\S]*?\}\s*\]?/g, '');
     cleanResponse = cleanResponse.replace(/\n{3,}/g, '\n\n').trim();
 
     if (!cleanResponse) {

src/utils/chatContentParser.ts

@@ -193,13 +193,16 @@ export function splitInlineToolCalls(text: string): ContentSegment[] {
         // Look for the next clearly non-JSON content boundary: a line starting with
         // a letter/header/markdown that isn't part of code content, or end of text
         const afterBlob = text.substring(endIdx);
-        const nextBoundary = afterBlob.search(/\n(?:#{1,3}\s|[A-Z][a-z]+ [a-z]|$|\n\n)/m);
-        if (nextBoundary > 0) {
-          skipEnd = endIdx + nextBoundary;
-        } else if (afterBlob.length < 5000) {
-          // Small remainder — likely all part of the leaked blob, consume it entirely
-          skipEnd = text.length;
+        // Find the next double-newline paragraph break — content before it is likely
+        // leaked code from the blob, content after it is likely the model's prose response.
+        const paraBreak = afterBlob.indexOf('\n\n');
+        if (paraBreak > 0) {
+          // Only extend to the paragraph break — preserve everything after it
+          skipEnd = endIdx + paraBreak;
         }
+        // If no paragraph break found, DON'T consume to end — the brace counter's endIdx
+        // is the best we have. Some content may leak, but that's better than swallowing
+        // the model's actual follow-up prose.
       }
       lastIndex = skipEnd;
       jsonRegex.lastIndex = lastIndex;